How to install ComfyUI and ComfyUI-Manager on a cloud GPU

Rent a GPU on Deploy, then open a terminal in JupyterLab or over SSH. Install ComfyUI from the official repository with the manual steps that README prints, enable ComfyUI-Manager the way that same README documents, and open the page through an SSH tunnel to 127.0.0.1 port 8188. Copy the output folder off the instance before you destroy it. Destroying the instance deletes its disk.

Which card to rent for a checkpoint is How to choose a rental GPU for ComfyUI. This page does not repeat that guide’s weight table. Checkpoint installs that already assume ComfyUI is running: How to run FLUX.1-dev in ComfyUI on a cloud GPU, How to run FLUX.1 Kontext in ComfyUI on a cloud GPU, and How to run Stable Diffusion 3.5 in ComfyUI on a cloud GPU. How published FLUX and SDXL weight sizes are read: How much VRAM do I need for FLUX and SDXL. A different image UI, from the AUTOMATIC1111 repository, is How to run AUTOMATIC1111 Stable Diffusion WebUI on a cloud GPU. The ComfyUI audio example’s Stable Audio Open section is How to run Stable Audio Open on a cloud GPU.

The ComfyUI LICENSE file begins with “GNU GENERAL PUBLIC LICENSE”, “Version 3, 29 June 2007” (read 2026-11-12). The ComfyUI-Manager repository’s GitHub license field on that day was GPL-3.0. Read the license before you redistribute the program.

The commands below are copied from the upstream files named in each step. They were not run on a CloudGPU instance for this page. Anything this page adds around them is marked.

What the README says about memory

The ComfyUI README feature list includes the clause “as low as 4GB vram + 8GB ram” (read 2026-11-12). That clause is the README’s memory figure. This page does not copy the rest of that sentence. The README does not say that 4GB is enough for a named checkpoint. Checkpoint figures stay on the guides linked above.

The README’s QA section links a wiki page titled “Which GPU should I buy for ComfyUI”. This page does not copy that wiki.

GET /v1/public/compute-capabilities does not list GPUs. It returns feature flags. The memory column is vram_gb from gpu-prices.json, updated_at 2026-11-12T21:04:40.726Z. GET /v1/public/gpu-catalog the same day (generated_at 2026-11-12T21:04:46.496Z) listed the same names and the same vram_gb values, plus Ascend 910B at 64, which is not in gpu-prices.json and is not listed here. The catalog response also includes an availability field. This page does not copy it. Every vram_gb below is 8 or higher, so every row is above the README’s 4GB clause as a number. This page does not mark a row as enough for a checkpoint. Which listing you can start is on Deploy.

Catalog name vram_gb
RTX 2080 8
RTX 3080 10
RTX 2080 Ti 11
RTX 3080 Ti 12
V100 16G 16
A30 24
RTX 3090 24
RTX 4090 24
RTX 4090D 24
RTX 4080 SUPER 32G 32
RTX 5090 32
RTX 5090D 32
V100 32G 32
A100 40G 40
A800 40G 40
A40 48
RTX 4090 48G 48
RTX 4090D 48G 48
RTX A6000 48
NVIDIA RTX PRO 5000 72G 72
RTX PRO 5000 72G 72
A100 80G 80
A800 80G 80
H800 80
RTX 6000D 83
H20 96G 96
RTX PRO 6000 96G 96
RTX PRO 6000 Max-Q 96
H20-3e 141G 141

The same catalog lists a template named ComfyUI. Its id is flux-1-dev, min_vram_gb is 24, and recommended_gpu is RTX 4090 48G. Those are template fields, not the README’s 4GB clause. This page’s commands are the manual install on Raw GPU. They are not a test of that template.

Manual install

The README’s manual-install section says “Git clone this repo” and does not print the git command. The repository URL on that page is https://github.com/comfyanonymous/ComfyUI. This page adds the clone. The README does not create a virtual environment in that section, so this page does not add one:

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI

The GitHub releases API on 2026-11-12 listed the latest tag as v0.45.0, published 2026-11-10T21:42:42Z. The clone above does not pin that tag.

The same section says Python 3.14 works but some custom nodes may have issues, and that Python 3.13 is very well supported. It says torch 2.7 is minimally supported and that a cu130 or above version of pytorch is required on Nvidia 20 series and above. The NVIDIA block prints this stable install:

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130

The README also prints a nightly command that uses the cu132 index. This page copies the stable command. The dependencies line, run in the ComfyUI folder:

pip install -r requirements.txt

requirements.txt on 2026-11-12 listed comfyui-frontend-package==1.59.1, comfyui-workflow-templates==0.11.107, and comfyui-embedded-docs==0.5.16 among its lines. It did not list comfyui_manager.

The README also prints a separate comfy-cli block, pip install comfy-cli and comfy install. This page follows the manual install, not that block.

ComfyUI-Manager

The ComfyUI README links ComfyUI-Manager on the manager-v4 branch. Its setup, copied:

pip install -r manager_requirements.txt

python main.py --enable-manager

manager_requirements.txt on 2026-11-12 contained one line: comfyui_manager==4.2.2.

main.py on master, read the same day, calls handle_comfyui_manager_unavailable() when --enable-manager is set and comfyui_manager cannot be imported. That function logs a pip command that installs -r the path of manager_requirements.txt, and then sets enable_manager to false. This page did not trigger that log.

The manager-v4 README’s notice says V4.0 changed the install to pip instead of git clone. Its installation section says that when installing the latest ComfyUI, Manager is installed as a dependency and manual installation is no longer necessary. It also prints a nightly manual install: clone https://github.com/Comfy-Org/ComfyUI-Manager into a temporary directory, not into ComfyUI/custom_nodes, then pip install . inside that clone. The core requirements.txt read above did not list comfyui_manager. This page does not treat the manager-v4 sentence and that requirements file as the same claim. The commands this page copies for Manager are the two from the ComfyUI README.

The manager-v4 “How To Use” section says to click the Manager button, then Install Custom Nodes or Install Models. It says the Install button installs an item, and Try Install is for a custom node whose installation information cannot be confirmed. It says the Fetch Updates menu retrieves update data for custom nodes locally, and that the actual update is the Update button in the Install Custom Nodes menu. Update All, in the Snapshot-Manager section, saves a snapshot of the current installation. The same README’s security table lists “Update ComfyUI” at the low level. The ComfyUI README does not print a git pull command. This page does not add one.

The manager-v4 Paths section says that in V4.0.3b4 and later, configuration files are under <USER_DIRECTORY>/__manager/. It says the user directory defaults to ComfyUI/user when ComfyUI is started without a user-directory option. This page prints both the pin comfyui_manager==4.2.2 and the README’s “V4.0.3b4 and later” sentence. It does not compare those version strings beyond printing them.

Listen address and SSH

comfy/cli_args.py on master defines --listen with default 127.0.0.1. The help text says that if --listen is provided without an argument, it defaults to 0.0.0.0,::. The same file defines --port with default 8188 and the help text “Set the listen port.” This page does not pass --listen, so the default stays 127.0.0.1. The launch command is the README’s Manager command above.

To keep that process after the SSH session closes, this page wraps it in tmux. The README does not print tmux:

tmux new -s comfyui
python main.py --enable-manager

The tunnel command is not in the ComfyUI README. It is the form in how to use SSH port forwarding on a cloud GPU:

ssh -N -L 8188:127.0.0.1:8188 -p <ssh-port> <ssh-user>@<ssh-host>

Then open http://127.0.0.1:8188 on your computer. ComfyUI is the UI. The README’s running section does not document a separate Gradio app.

Model folders

The manual-install section says to put SD checkpoints (the huge ckpt/safetensors files) in models/checkpoints, and to put your VAE in models/vae. extra_model_paths.yaml.example lists further folder keys, including loras, text_encoders, diffusion_models, clip_vision, embeddings, controlnet, upscale_models, and custom_nodes. The example says to rename that file to extra_model_paths.yaml. --extra-model-paths-config in cli_args.py loads one or more of those files. This page does not download a checkpoint. The checkpoint guides linked above name their own files.

When it fails

“Torch not compiled with CUDA enabled”. The README’s NVIDIA troubleshooting says to uninstall torch with pip uninstall torch and install it again with the cu130 command above.

CUDA out of memory. cli_args.py says --lowvram does nothing if dynamic vram is enabled, and that if dynamic vram is not being used this option makes the text encoders run on the CPU. It says --novram is “When lowvram isn’t enough.” Those help strings are not a measured cutoff. If the process raises that error, use how to fix CUDA out of memory on a cloud GPU.

“CUDA error: no kernel image is available for execution on the device” on an RTX 5090. The README says a cu130 or above pytorch is required on Nvidia 20 series and above. It does not name RTX 5090. The portable section says the cu126 portable is for Nvidia 10 series and older and says not to use that portable on 20 series and above. The RTX 5090 guide says that card is compute capability 12.0 (sm_120) and needs a PyTorch build for CUDA 12.8 or newer. This page did not re-test the cu130 command on an RTX 5090. Catalog RTX 5090 and RTX 5090D are 32, which is above the README’s 4GB clause as a number. That comparison is the memory number only.

Save the images, then destroy

folder_paths.py sets output_directory to os.path.join(base_path, "output"). cli_args.py says --output-directory sets the output directory and overrides --base-directory. Copy that directory off before you destroy the instance. Save your work before you destroy a cloud GPU is the checklist. The terms say data on local instance storage is permanently deleted when the instance is destroyed.

The pricing page states the billing rule in these words:

Per minute, rounded up, with a 1-minute minimum. One hour is pre-authorised while an instance runs, and unused minutes go back to your balance when you destroy the instance. There is no subscription and no minimum term.

Prices are not copied here.

Frequently asked questions

How much VRAM does the ComfyUI README publish?

The ComfyUI README feature list, read 2026-11-12, includes the clause as low as 4GB vram + 8GB ram. It does not say that figure is enough for a named checkpoint. Every vram_gb in https://cloudgpu.app/gpu-prices.json (updated_at 2026-11-12T21:04:40.726Z) is 8 or higher, so every row is above that 4GB clause as a number. Which card to rent for a checkpoint is https://cloudgpu.app/guides/how-to-choose-a-rental-gpu-for-comfyui. Which listing you can start is on https://cloudgpu.app/deploy.

Which address and port does ComfyUI listen on by default?

comfy/cli_args.py on master defines --listen with default 127.0.0.1 and --port with default 8188. The help text says that --listen without an argument defaults to 0.0.0.0,::. This page does not pass --listen. The SSH tunnel forwards 127.0.0.1 port 8188. The launch command the ComfyUI README prints for Manager is python main.py --enable-manager.

How do I install or update a custom node with ComfyUI-Manager?

The manager-v4 README says to click the Manager button, then Install Custom Nodes, then Install or Try Install. It says Fetch Updates retrieves update data locally and the Update button applies it. Update All saves a snapshot. The ComfyUI README enables Manager with python main.py --enable-manager after pip install -r manager_requirements.txt. manager_requirements.txt pins comfyui_manager==4.2.2. The ComfyUI README does not print a git pull command.

Do generated images stay on the instance after I destroy it?

No. folder_paths.py joins the base path with the folder name output. cli_args.py says --output-directory overrides that directory. Copy the directory off before you destroy the instance. The terms say data on local instance storage is permanently deleted when the instance is destroyed.

Sources (read 2026-11-12, Asia/Shanghai)

Cards you can start are on the Deploy page. Prices are on https://cloudgpu.app/pricing and are not repeated here.

Last updated: 2026-11-12 (Asia/Shanghai).